Measurement, Modeling, and Analysis of a Peer-to-Peer File sharing Workload

1. Measurement, Modeling, and Analysis of a Peer-to-Peer File sharing Workload Krishna P. Gummadi, Richard J. Dunn, Stefan Saroiu, Steven D. Gribble, Henry M. Levy, John Zahorjan

2. Outline • Motivation • Goals • Approach • Analysis of Users • Analysis of Objects • Kazaa is not Zipf • Exploiting Locality • Conclusion

3. Motivation • Dramatic shift of Internet traffic from WWW to multimedia file sharing • March 2000 study found that bandwidth consumed by Napster was greater than HTTP • On the UDUB campus, peer-to-peer file sharing consumed 43%, WWW traffic 14% • Multimedia file sharing dominates now, and will dominate Internet of the future

4. Goals • To understand the fundamental properties of multimedia file-sharing systems • To explore the forces driving P2P file-sharing workloads • To demonstrate that opportunity exists to optimize performance in current file-sharing workloads

5. Approach • Analyze a 200-day trace of Kazaa traffic at the University of Washington • Over 60,000 faculty, students, and staff • 20 TBs of incoming data (1.6 million requests) • Long enough to observe seasonal variations • Derive a model of this multimedia traffic • Use simulation to quantify the potential to improve performance of file-sharing

6. Analysis of Users • Kazaa users are patient • In the WWW, users expect instant results • The Web is an interactive system, whereas Kazaa is a batch-mode delivery system

7. Analysis of Users (continued..) • Users slow down as they age • Older clients consume fewer bytes than newer clients • Due to attrition (clients leaving the system forever) and older clients having slower request rates

8. User Summary • New clients generate most of the load in Kazaa • Older clients consume fewer bytes as they age • This is because of attrition: clients leave the system permanently as they grow older. • Older clients also tend to interact with the system at a constant rate, but ask for less during each interaction.

9. Analysis of Objects • Small objects take up the least of the bandwidth • However, most requests are for small objects

10. Analysis of Objects (continued..) • Majority of requests are for small objects • Majority of bytes transferred are due to the largest objects

11. Analysis of Objects (continued..) • Crucial difference (Web/multimedia): • Multimedia objects are immutable • Kazaa clients fetch objects at most once • 94% an object is requested at most once • Popularity of Kazaa objects is often short-lived • Most popular objects tend to be recently born • Most requests are for old objects • Large objects requested tend to be older than small objects

12. Kazaa is not Zipf • Zipf’s law: popularity of ith-most popular object is proportional to i-a • Distribution looks linear when plotted on a log-log scale

13. Kazaa is not Zipf (continued..) • The most popular objects are requested much less, while objects down the tail show elevated number of requests.

14. Exploiting Locality • Exploitation of locality in file-sharing • To decrease external bandwidth usage • There is a tremendous amount of untapped locality in the Kazaa workload • Used a proxy cache at the organizational border, guaranteeing that every object is downloaded into the organization at most once • Additional requests satisfied without consuming external bandwidth

15. Exploiting Locality (continued..) • 68% byte hit rate for large objects (22.3 TB saved) • 37% byte hit rate for small objects (1.5 TB saved)

16. Conclusion • Client/object births drive P2P file-sharing • Changes to objects drive the Web • Fetch-at-most-once causes distribution of objects to deviate substantially from Zipf • There is significant locality in Kazaa • Opportunity for caching to reduce wide-area bandwidth consumption

17. Any questions?